feat(ai): AI backbone + inline chat for code review#363
Merged
Conversation
Introduces @plannotator/ai — a pluggable provider layer that powers AI features (inline chat, plan Q&A, code review assistance) across all Plannotator surfaces. - ProviderRegistry: class-based, multi-instance provider management - SessionManager: in-memory session tracking with eviction, ID aliasing - Context builders: system prompts and fork preambles for all three modes - SSE streaming endpoints: session create/fork, query, abort, list - Claude Agent SDK provider: fork from parent sessions, read-only tools, per-query abort, cached SDK import, concurrent query guard - 34 passing tests covering all layers For provenance purposes, this commit was AI assisted.
- Concurrent query race: reject with session_busy error instead of
racing; add query generation counter so stale finally blocks don't
clobber newer query state
- Text dropped before tool_use: mapSDKMessage now returns AIMessage[]
and flushes accumulated text before each tool_use block
- defaultProvider returns instance ID: ProviderRegistry.getDefault()
returns { id, provider } so capabilities endpoint sends the correct
key for multi-instance setups
- Fork preamble re-sent on error: use _firstQuerySent boolean instead
of _resolvedId to guard preamble injection
- Dead abortController params: remove unused abortController from
CreateSessionOptions and AIProvider.resumeSession
- Unused context in memory: remove context from SessionConfig and
baseConfig — already consumed to build systemPrompt/forkPreamble
- Fix tool_result mapping: SDK wraps tool results in SDKUserMessage
(type "user"), not a "tool_result" type
For provenance purposes, this commit was AI assisted.
Add "Ask AI" to the code review toolbar — select lines, ask questions about agent-written code, get streaming responses. Conversation persists across file switches with inline markers on the diff showing where questions were asked. - useAIChat hook: SSE streaming, lazy session creation, multi-turn - AskAIInput: question input with per-selection AI history - InlineAIMarker: compact sparkle markers on diff lines via @pierre - AITab: sidebar conversation grouped by file with collapsible sections - SparklesIcon: shared 3-star sparkle from automations branch - ReviewPanel: AI tab with message count badge, file group flash highlight - Server: mount AI endpoints in review server with graceful fallback - Support three question scopes: line, file, and general - Toolbar ergonomics: typed text submits as AI question, no double popover For provenance purposes, this commit was AI assisted.
- X button closes entire toolbar in both modes (no ghost popover) - Comment/Ask AI toggle is inline in the action row, same position in both modes (left side, mirrored layout) - AI history appears below the input, not above - File group flash highlight uses --muted to match file tree hover For provenance purposes, this commit was AI assisted.
Resolves import conflict in packages/server/review.ts — main renamed getGhUser to getUser and added GitLab MR helpers (prRefFromMetadata, getDisplayRepo, getMRLabel, getMRNumberLabel). Kept both main's updated PR imports and our AI endpoint import. For provenance purposes, this commit was AI assisted.
Visual consistency: - AI tab file group headers match FileTree pattern (sizing, hover, badges) - Q&A bubbles match annotation card padding and border-radius - Empty state uses circle-bg icon (same as annotations) - General input matches search input styling - Removed redundant CSS classes, moved to Tailwind Shared primitives: - Extract formatTimestamp to shared utility (was duplicated) - CountBadge component for consistent badge styling - CopyButton component with hover-reveal + "Copied" flash Affordances: - Copy button on AI responses (hover-reveal) - Send button on general input with keyboard shortcut hint - Timestamps on Q&A pairs - Keyboard shortcut title on Ask button Markdown rendering: - Add marked dependency for full block-level markdown in AI responses - renderMarkdown utility using marked + DOMPurify - Scoped .ai-markdown CSS for headings, lists, code blocks, blockquotes For provenance purposes, this commit was AI assisted.
The inline AI agent now inherits the user's existing Claude Code permission rules via settingSources: ['user', 'project']. Tools already approved in ~/.claude/settings.json auto-approve silently. For tools not covered by existing rules, the SDK emits control_request messages. These are mapped to a new AIPermissionRequestMessage type and surfaced as inline approval cards in the AI tab — showing the tool name, command/input, and Allow/Deny buttons. Backend: - Map control_request to AIPermissionRequestMessage in provider - Store Query object on session for sending control_response back - respondToPermission() method on AISession sends allow/deny - POST /api/ai/permission endpoint for frontend → backend decisions - Add settingSources to ClaudeSDKQueryOptions Frontend: - PermissionCard component with Allow/Deny buttons + resolved state - useAIChat tracks permissionRequests array - respondToPermission() POSTs decision to server - AITab renders pending permission cards inline in conversation For provenance purposes, this commit was AI assisted.
Streaming fixes: - State updates now create new object references via prev.map() instead of mutating objects inside the array — fixes React not re-rendering the sidebar during streaming - Complete assistant text messages only used as fallback when no streaming deltas received (prevents duplicate text) - Result text used as fallback when response is empty (agent spent all turns on tools) Config: - maxTurns raised to 99 (was 3 — agent ran out of turns reading files and never got to write its answer) Popover → sidebar linking: - Clicking AI history items in comment popover now passes questionId through to handleViewAIResponse, triggering the file group flash highlight in the sidebar (same behavior as clicking inline markers) Debug logging (temporary): - Server-side: [AI STREAM] logs every message the provider yields - Client-side: [AI SSE] logs every message received in the browser For provenance purposes, this commit was AI assisted.
From the three-agent code review analysis: Quality: - Rename AIMessage → AIChatEntry in review-editor to avoid collision with the wire-protocol AIMessage from packages/ai - Type permissionMode as union (not string) in SessionConfig + QueryOptions - Type _activeQuery with narrow interface instead of any - Make toolUseId optional on AIToolResultMessage (SDK doesn't populate it) - Remove redundant sessionIdRef double-write in useAIChat - Fix handleAskGeneral useCallback dep: [aiChat] → [aiChat.ask] Performance: - Wrap QAPair in React.memo to prevent re-renders on sibling updates - Memoize renderMarkdown output via useMemo (avoid re-parsing on every token) - Fix auto-scroll useEffect: depend on messages.length not messages ref (prevents querySelectorAll on every streaming token) Cleanup: - Remove debug console.log from SSE endpoint and browser parser - Extract isMac/submitHint to shared utils/platform.ts - Extract generateId to shared utils/generateId.ts Net: -2 lines. No behavior changes. For provenance purposes, this commit was AI assisted.
backnotprop
force-pushed
the
feat/ai-backbone
branch
from
431c3ab to
7f96328
Compare
- Use `import type` where only types are consumed (endpoints, CountBadge, CopyButton, renderMarkdown) - Remove unused AIContext import from session-manager - Template literal instead of string concat in context.ts truncate() - Add type="button" + aria-hidden on CopyButton SVGs For provenance purposes, this commit was AI assisted.
- Add CodexSDKProvider implementing AIProvider interface - Maps Codex ThreadEvents to AIMessage types - Delta tracking for streaming text from cumulative item.updated events - Supports resume, abort, sandbox modes - Factory self-registration pattern matching Claude provider - Enhance capabilities endpoint with provider models + metadata - Register both providers in review server (graceful degradation) - Add CodexSDKConfig type, models list (GPT-5.4 family) - Add reasoningEffort to CreateSessionOptions (Codex only) - Extract DiffFile interface to shared types.ts - Update CLAUDE.md with build order documentation - 54 tests passing (20+ new for Codex event mapping) For provenance purposes, this commit was AI assisted.
- Settings → AI tab: provider cards with per-provider model dropdowns - Inline config bar in AI chat sidebar (provider, model, reasoning effort) - Provider icons: Claude SVG + Codex PNG from marketing assets - Cookie-based persistence for provider + per-provider model preferences - useAIChat accepts providerId, model, reasoningEffort from props - Session reset on provider/model switch (next question starts fresh) - Reasoning effort selector (Codex only: Low/Medium/High/Max) - Update code review docs with AI provider documentation For provenance purposes, this commit was AI assisted.
…on, codex-sdk dep - Add missing selectedReasoningEffort + onReasoningEffortChange props to populated-state AIConfigBar (was only on empty-state, causing crash) - Disable GeneralInput while streaming to prevent concurrent queries - Thread cwd through CreateSessionOptions so AI sessions use the correct working directory (worktree-aware via getCwd resolver) - Add @openai/codex-sdk as a real dependency (was only an optional peer) - Eagerly verify codex-sdk importability at startup to avoid advertising a broken provider For provenance purposes, this commit was AI assisted.
- Replace native <select> dropdowns with custom popover menus that show provider icons, model names, and checkmarks on active selection - Remove disabled gear placeholder button - Auto-grow textarea upward on multi-line input (capped at 6 lines) - Push tab icons to far right in sidebar header (w-full on flex container) - Add centered divider between file-scoped and general questions For provenance purposes, this commit was AI assisted.
backnotprop
changed the title
- Endpoint now checks capabilities.fork before calling forkSession(), falling back to createSession() for providers that can't truly fork - Codex forkSession() now throws as the interface contract requires (was silently faking a fork with no history — misleading UX) - Fixed permissionMode JSDoc: said "plan" but code defaults to "default" For provenance purposes, this commit was AI assisted.
Count now appears inline to the left of the X button instead of squeezing the input from outside. Dynamic right padding accommodates the count + clear button when a query is active. For provenance purposes, this commit was AI assisted.
…d/ listing For provenance purposes, this commit was AI assisted.
- Claude provider now uses { preset: "claude_code", append } instead of
a raw system prompt, so inline chat inherits Claude Code's full
built-in capabilities
- Stripped custom role/guidelines preamble from context builders — just
emit the review context (diff, plan, document)
- handleAskAI now extracts selected code via extractLinesFromPatch and
sends it with the query, fixing hallucinations on large PRs where the
file's diff was truncated out of the system prompt
For provenance purposes, this commit was AI assisted.
Both @anthropic-ai/claude-agent-sdk and @openai/codex-sdk are now direct imports (no string indirection) so bun build --compile bundles them into the binary. Users don't need to install SDKs separately. Removed peer dependency declarations from packages/ai/package.json — meaningless for a compiled binary. For provenance purposes, this commit was AI assisted.
- Created packages/ui/utils/platform.ts as the single source for isMac, modKey, altKey, submitHint, isWindows - Consolidated inline navigator.platform checks from 6 files - Renamed renderMarkdown → renderChatMarkdown (distinct from plan editor) - Renamed formatTimestamp → formatRelativeTime (clearer purpose) - Added file-level comments to InlineAIMarker, PermissionCard, generateId For provenance purposes, this commit was AI assisted.
New docs page at guides/ai-features covering providers, models, configuration, session behavior, permissions, and available settings. AISettingsTab now links to the docs in both empty and populated states. For provenance purposes, this commit was AI assisted.
stk-code
pushed a commit
to stk-code/plannotator
that referenced
this pull request
Mar 31, 2026
* feat(ai): add provider-agnostic AI backbone for inline chat Introduces @plannotator/ai — a pluggable provider layer that powers AI features (inline chat, plan Q&A, code review assistance) across all Plannotator surfaces. - ProviderRegistry: class-based, multi-instance provider management - SessionManager: in-memory session tracking with eviction, ID aliasing - Context builders: system prompts and fork preambles for all three modes - SSE streaming endpoints: session create/fork, query, abort, list - Claude Agent SDK provider: fork from parent sessions, read-only tools, per-query abort, cached SDK import, concurrent query guard - 34 passing tests covering all layers For provenance purposes, this commit was AI assisted. * fix(ai): address code review findings from PR backnotprop#363 - Concurrent query race: reject with session_busy error instead of racing; add query generation counter so stale finally blocks don't clobber newer query state - Text dropped before tool_use: mapSDKMessage now returns AIMessage[] and flushes accumulated text before each tool_use block - defaultProvider returns instance ID: ProviderRegistry.getDefault() returns { id, provider } so capabilities endpoint sends the correct key for multi-instance setups - Fork preamble re-sent on error: use _firstQuerySent boolean instead of _resolvedId to guard preamble injection - Dead abortController params: remove unused abortController from CreateSessionOptions and AIProvider.resumeSession - Unused context in memory: remove context from SessionConfig and baseConfig — already consumed to build systemPrompt/forkPreamble - Fix tool_result mapping: SDK wraps tool results in SDKUserMessage (type "user"), not a "tool_result" type For provenance purposes, this commit was AI assisted. * feat(ai): inline AI chat on code review diffs Add "Ask AI" to the code review toolbar — select lines, ask questions about agent-written code, get streaming responses. Conversation persists across file switches with inline markers on the diff showing where questions were asked. - useAIChat hook: SSE streaming, lazy session creation, multi-turn - AskAIInput: question input with per-selection AI history - InlineAIMarker: compact sparkle markers on diff lines via @pierre - AITab: sidebar conversation grouped by file with collapsible sections - SparklesIcon: shared 3-star sparkle from automations branch - ReviewPanel: AI tab with message count badge, file group flash highlight - Server: mount AI endpoints in review server with graceful fallback - Support three question scopes: line, file, and general - Toolbar ergonomics: typed text submits as AI question, no double popover For provenance purposes, this commit was AI assisted. * fix(ai): toolbar popover ergonomics and visual consistency - X button closes entire toolbar in both modes (no ghost popover) - Comment/Ask AI toggle is inline in the action row, same position in both modes (left side, mirrored layout) - AI history appears below the input, not above - File group flash highlight uses --muted to match file tree hover For provenance purposes, this commit was AI assisted. * fix(ai): visual consistency, shared primitives, markdown rendering Visual consistency: - AI tab file group headers match FileTree pattern (sizing, hover, badges) - Q&A bubbles match annotation card padding and border-radius - Empty state uses circle-bg icon (same as annotations) - General input matches search input styling - Removed redundant CSS classes, moved to Tailwind Shared primitives: - Extract formatTimestamp to shared utility (was duplicated) - CountBadge component for consistent badge styling - CopyButton component with hover-reveal + "Copied" flash Affordances: - Copy button on AI responses (hover-reveal) - Send button on general input with keyboard shortcut hint - Timestamps on Q&A pairs - Keyboard shortcut title on Ask button Markdown rendering: - Add marked dependency for full block-level markdown in AI responses - renderMarkdown utility using marked + DOMPurify - Scoped .ai-markdown CSS for headings, lists, code blocks, blockquotes For provenance purposes, this commit was AI assisted. * feat(ai): permission inheritance + inline approval UI The inline AI agent now inherits the user's existing Claude Code permission rules via settingSources: ['user', 'project']. Tools already approved in ~/.claude/settings.json auto-approve silently. For tools not covered by existing rules, the SDK emits control_request messages. These are mapped to a new AIPermissionRequestMessage type and surfaced as inline approval cards in the AI tab — showing the tool name, command/input, and Allow/Deny buttons. Backend: - Map control_request to AIPermissionRequestMessage in provider - Store Query object on session for sending control_response back - respondToPermission() method on AISession sends allow/deny - POST /api/ai/permission endpoint for frontend → backend decisions - Add settingSources to ClaudeSDKQueryOptions Frontend: - PermissionCard component with Allow/Deny buttons + resolved state - useAIChat tracks permissionRequests array - respondToPermission() POSTs decision to server - AITab renders pending permission cards inline in conversation For provenance purposes, this commit was AI assisted. * fix(ai): streaming reactivity, maxTurns, popover→sidebar linking Streaming fixes: - State updates now create new object references via prev.map() instead of mutating objects inside the array — fixes React not re-rendering the sidebar during streaming - Complete assistant text messages only used as fallback when no streaming deltas received (prevents duplicate text) - Result text used as fallback when response is empty (agent spent all turns on tools) Config: - maxTurns raised to 99 (was 3 — agent ran out of turns reading files and never got to write its answer) Popover → sidebar linking: - Clicking AI history items in comment popover now passes questionId through to handleViewAIResponse, triggering the file group flash highlight in the sidebar (same behavior as clicking inline markers) Debug logging (temporary): - Server-side: [AI STREAM] logs every message the provider yields - Client-side: [AI SSE] logs every message received in the browser For provenance purposes, this commit was AI assisted. * refactor(ai): simplification pass — types, perf, cleanup From the three-agent code review analysis: Quality: - Rename AIMessage → AIChatEntry in review-editor to avoid collision with the wire-protocol AIMessage from packages/ai - Type permissionMode as union (not string) in SessionConfig + QueryOptions - Type _activeQuery with narrow interface instead of any - Make toolUseId optional on AIToolResultMessage (SDK doesn't populate it) - Remove redundant sessionIdRef double-write in useAIChat - Fix handleAskGeneral useCallback dep: [aiChat] → [aiChat.ask] Performance: - Wrap QAPair in React.memo to prevent re-renders on sibling updates - Memoize renderMarkdown output via useMemo (avoid re-parsing on every token) - Fix auto-scroll useEffect: depend on messages.length not messages ref (prevents querySelectorAll on every streaming token) Cleanup: - Remove debug console.log from SSE endpoint and browser parser - Extract isMac/submitHint to shared utils/platform.ts - Extract generateId to shared utils/generateId.ts Net: -2 lines. No behavior changes. For provenance purposes, this commit was AI assisted. * chore: biome lint fixes — import types, button a11y, unused import - Use `import type` where only types are consumed (endpoints, CountBadge, CopyButton, renderMarkdown) - Remove unused AIContext import from session-manager - Template literal instead of string concat in context.ts truncate() - Add type="button" + aria-hidden on CopyButton SVGs For provenance purposes, this commit was AI assisted. * feat(ai): Codex SDK provider + multi-provider backend - Add CodexSDKProvider implementing AIProvider interface - Maps Codex ThreadEvents to AIMessage types - Delta tracking for streaming text from cumulative item.updated events - Supports resume, abort, sandbox modes - Factory self-registration pattern matching Claude provider - Enhance capabilities endpoint with provider models + metadata - Register both providers in review server (graceful degradation) - Add CodexSDKConfig type, models list (GPT-5.4 family) - Add reasoningEffort to CreateSessionOptions (Codex only) - Extract DiffFile interface to shared types.ts - Update CLAUDE.md with build order documentation - 54 tests passing (20+ new for Codex event mapping) For provenance purposes, this commit was AI assisted. * feat(ai): provider settings UI + inline config bar - Settings → AI tab: provider cards with per-provider model dropdowns - Inline config bar in AI chat sidebar (provider, model, reasoning effort) - Provider icons: Claude SVG + Codex PNG from marketing assets - Cookie-based persistence for provider + per-provider model preferences - useAIChat accepts providerId, model, reasoningEffort from props - Session reset on provider/model switch (next question starts fresh) - Reasoning effort selector (Codex only: Low/Medium/High/Max) - Update code review docs with AI provider documentation For provenance purposes, this commit was AI assisted. * fix(ai): review fixes — reasoning props, input blocking, cwd resolution, codex-sdk dep - Add missing selectedReasoningEffort + onReasoningEffortChange props to populated-state AIConfigBar (was only on empty-state, causing crash) - Disable GeneralInput while streaming to prevent concurrent queries - Thread cwd through CreateSessionOptions so AI sessions use the correct working directory (worktree-aware via getCwd resolver) - Add @openai/codex-sdk as a real dependency (was only an optional peer) - Eagerly verify codex-sdk importability at startup to avoid advertising a broken provider For provenance purposes, this commit was AI assisted. * fix(ai): polish config bar, input, and sidebar header - Replace native <select> dropdowns with custom popover menus that show provider icons, model names, and checkmarks on active selection - Remove disabled gear placeholder button - Auto-grow textarea upward on multi-line input (capped at 6 lines) - Push tab icons to far right in sidebar header (w-full on flex container) - Add centered divider between file-scoped and general questions For provenance purposes, this commit was AI assisted. * fix(ai): honest fork capability + permissionMode doc accuracy - Endpoint now checks capabilities.fork before calling forkSession(), falling back to createSession() for providers that can't truly fork - Codex forkSession() now throws as the interface contract requires (was silently faking a fork with no history — misleading UX) - Fixed permissionMode JSDoc: said "plan" but code defaults to "default" For provenance purposes, this commit was AI assisted. * fix(review): move search result count inside input field Count now appears inline to the left of the X button instead of squeezing the input from outside. Dynamic right padding accommodates the count + clear button when a query is active. For provenance purposes, this commit was AI assisted. * fix(ai): use claude_code preset, pass selected code to queries - Claude provider now uses { preset: "claude_code", append } instead of a raw system prompt, so inline chat inherits Claude Code's full built-in capabilities - Stripped custom role/guidelines preamble from context builders — just emit the review context (diff, plan, document) - handleAskAI now extracts selected code via extractLinesFromPatch and sends it with the query, fixing hallucinations on large PRs where the file's diff was truncated out of the system prompt For provenance purposes, this commit was AI assisted. * fix(ai): bundle both SDKs, remove peer dependencies Both @anthropic-ai/claude-agent-sdk and @openai/codex-sdk are now direct imports (no string indirection) so bun build --compile bundles them into the binary. Users don't need to install SDKs separately. Removed peer dependency declarations from packages/ai/package.json — meaningless for a compiled binary. For provenance purposes, this commit was AI assisted. * refactor: canonical platform.ts, file renames, file-level comments - Created packages/ui/utils/platform.ts as the single source for isMac, modKey, altKey, submitHint, isWindows - Consolidated inline navigator.platform checks from 6 files - Renamed renderMarkdown → renderChatMarkdown (distinct from plan editor) - Renamed formatTimestamp → formatRelativeTime (clearer purpose) - Added file-level comments to InlineAIMarker, PermissionCard, generateId For provenance purposes, this commit was AI assisted. * docs(ai): add AI features guide + link from settings New docs page at guides/ai-features covering providers, models, configuration, session behavior, permissions, and available settings. AISettingsTab now links to the docs in both empty and populated states. For provenance purposes, this commit was AI assisted.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
@plannotator/ai— a pluggable provider layer powering AI features across all Plannotator surfacesWhat the user sees
Architecture
Key design decisions
Dependencies added
@anthropic-ai/claude-agent-sdk^0.2.81@openai/codex-sdk^0.116.0Test plan
bun test packages/ai/)apps/review build→build:hook→ compiled binary